[build] Use static linking for CUDA build and allow dynamic AVX dispatch

Also add caching to speedup builds.

Author: Breakthrough

.github/workflows/build-windows-cuda.yml

@@ -24,32 +24,59 @@ jobs:
       - name: 📥 Checkout code
         uses: actions/checkout@v4
 
-      - name: 🏗️ Install Build Dependencies
-        run: |
-          echo "Installing CMake and Ninja for building..."
-          choco install -y cmake ninja
+      #- name: 🏗️ Install Build Dependencies
+      #  run: |
+      #    echo "Installing CMake and Ninja for building..."
+      #    choco install -y cmake ninja
 
       - name: 🐍 Set up Python
         uses: actions/setup-python@v5
         with:
-          python-version: '3.13'
+          python-version: '3.12'
 
       - name: 📦 Install Python Build Requirements
         run: |
           python -m pip install --upgrade pip
           pip install numpy wheel
 
+      - name: Setup MSBuild.exe
+        uses: microsoft/setup-msbuild@v1.1
+
+      - name: Setup NASM
+        uses: ilammy/setup-nasm@v1
+
+      - name: 🏗️ Setup Visual Studio Developer Command Prompt
+        run: |
+          $vcvars_path = "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          $lines = Invoke-Expression "& `"$vcvars_path`" && set"
+          $lines | ForEach-Object {
+            if ($_ -match "=") {
+              $name, $value = $_.Split("=", 2)
+              echo "$name=$value" | Out-File -FilePath $env:GITHUB_ENV -Append
+            }
+          }
+        shell: pwsh
+
+      - name: Cache CUDA Toolkit Installer
+        id: cache-cuda-installer
+        uses: actions/cache@v3
+        with:
+          path: ./.cache/cuda_installer.exe
+          key: cuda-installer-12.1.1
+
       - name: 🔧 Install NVIDIA CUDA Toolkit
         run: |
-          echo "Downloading CUDA Toolkit..."
-          $cuda_installer_url = "https://developer.download.nvidia.com/compute/cuda/12.1.1/network_installers/cuda_12.1.1_windows_network.exe"
-          # Use the runner's temp directory, which is guaranteed to exist.
-          $installer_path = Join-Path $env:RUNNER_TEMP "cuda_installer.exe"
-          curl -L -o $installer_path $cuda_installer_url
+          $installer_path = "./.cache/cuda_installer.exe"
+          if (-not (Test-Path $installer_path)) {
+            echo "Downloading CUDA Toolkit..."
+            $cuda_installer_url = "https://developer.download.nvidia.com/compute/cuda/12.1.1/network_installers/cuda_12.1.1_windows_network.exe"
+            New-Item -Path (Split-Path $installer_path) -ItemType Directory -Force
+            curl -L -o $installer_path $cuda_installer_url
+          } else {
+            echo "CUDA Toolkit installer found in cache."
+          }
 
           echo "Installing CUDA Toolkit silently..."
-          # Use Start-Process for robust execution in PowerShell.
-          # The arguments must be passed as a single string.
           $arguments = "-s nvcc_12.1 cudart_12.1"
           Start-Process -FilePath $installer_path -ArgumentList $arguments -Wait -NoNewWindow
 
@@ -60,20 +87,36 @@ jobs:
           echo "$cuda_path\lib\x64" | Out-File -FilePath $env:GITHUB_PATH -Append
         shell: pwsh
 
+      - name: Cache opencv-python repository
+        id: cache-opencv-python
+        uses: actions/cache@v3
+      # TODO: Need to pin this to a version.
+        with:
+          path: opencv-python
+          key: opencv-python-master
+
       - name: Clone opencv-python repository
+        if: steps.cache-opencv-python.outputs.cache-hit != 'true'
         run: |
           git clone --depth 1 --recursive https://github.com/opencv/opencv-python.git
-          pip install numpy packaging scikit-build
+
+      - name: Install opencv-python dependencies
+        run: |
+          pip install numpy packaging scikit-build cmake==3.24.2
 
         # - CUDA_ARCH_BIN: Specifies the CUDA architectures to build for.
         # - CUDA_ARCH_PTX: Specifies the virtual architecture for forward compatibility.
         # TODO: Set correct PTX versions below.
         # TODO: Validate the built wheel can be installed and run.
+        # TODO: Slim down the modules we build to only the ones that DVR-Scan needs to run:
+        # https://docs.opencv.org/4.x/db/d05/tutorial_config_reference.html#tutorial_config_reference_general_modules
       - name: 🛠️ Build opencv-python with CUDA
         run: |
           cd opencv-python
-          $CMAKE_ARGS = '-D CMAKE_BUILD_TYPE=RELEASE -D WITH_CUDA=ON -D CUDA_ARCH_BIN="6.0;6.1;7.0;7.5" -D CUDA_ARCH_PTX=7.5 -D OPENCV_ENABLE_NONFREE=ON -D BUILD_TESTS=OFF -D BUILD_PERF_TESTS=OFF -D BUILD_EXAMPLES=OFF'
+          $CMAKE_ARGS = '-D CMAKE_BUILD_TYPE=RELEASE -D BUILD_SHARED_LIBS=OFF -D WITH_CUDA=ON -D CUDA_ARCH_BIN="6.0;6.1;7.0;7.5" -D CUDA_ARCH_PTX=7.5 -D OPENCV_ENABLE_NONFREE=ON -D ENABLE_LTO=ON -D CPU_DISPATCH="AVX,AVX2" -D BUILD_TESTS=OFF -D BUILD_PERF_TESTS=OFF -D BUILD_EXAMPLES=OFF -D OPENCV_EXTRA_MODULES_PATH=opencv_contrib/modules'
+          $ENABLE_CONTRIB = 1
           pip wheel . --verbose
+        shell: pwsh
 
       - name: 📤 Upload Artifact
         uses: actions/upload-artifact@v4.6.0